NIVQRA · Authority & Control LayerGOV-01 · AI agent governance

AI agent governance

Agents are starting to act on behalf of organisations — buying tools, commissioning research, moving money. Governance is what turns that from an experiment into something an operating committee can sign off: explicit authority, enforcement that does not depend on a model's judgement, a person accountable for every exception, and the ability to withdraw permission at any moment.

Sandbox only. No custody, no customer funds and no live payment execution.

Five controls that make agent authority governable

01

Mandate — written authority, not implied trust

Every agent operates under an explicit mandate: a monthly ceiling, a per-transaction limit, permitted and prohibited categories, and the threshold above which a human must sign. Governance starts by writing the authority down in a form a system can enforce, rather than relying on a prompt or a policy document nobody can execute.

02

Deterministic evaluation — fixed rules, not model judgement

No language model decides whether an action is allowed. Each intent is evaluated by ordinary application logic in a fixed order — category restrictions, per-transaction ceiling, remaining monthly authority, approval threshold, new-counterparty rule — and every check returns a readable reason. The same input always produces the same decision, which is what makes a control auditable.

03

Named accountability — a person behind every exception

Anything outside the mandate routes to a named approver with the decision rights for that scope, not to an anonymous queue. Governance fails when an exception can be cleared by whoever happens to be online; it holds when the record shows which human accepted the risk.

04

Evidence — a chain that can be read after the fact

Each action produces a sequential record: what was attempted, which clause applied, what the outcome was, who was accountable and when. Evidence is generated as the decision is made, not reconstructed from logs during a review.

05

Revocation — authority you can withdraw immediately

Delegated authority is only safe if it can be taken back. Any mandate can be suspended or revoked in one action, and the agent's next attempt is refused rather than queued. Governance is the ability to stop, not only to observe.

A maturity path for delegated authority

  1. 01ObserveThe agent proposes; nothing executes. Used to learn the shape of real demand before granting authority.
  2. 02ApproveEvery action requires a human signature. Slow by design, appropriate for new or high-risk scopes.
  3. 03BoundAction inside written limits clears automatically; exceptions escalate. The working state for most agents.
  4. 04AutonomizeWider limits earned by demonstrated behaviour, still bounded, still revocable, still fully recorded.

In the NIVQRA demonstration workspace, the fictional operator Northstar Systems runs agents at each of these stages at once — a research agent under observation, a procurement agent operating inside written limits, and a growth agent whose prohibited-category attempts are refused outright.

Common questions

What is AI agent governance?
AI agent governance is the set of controls that determine what an autonomous agent may access, spend, execute and approve — and the evidence that proves what it actually did. In practice it combines written mandates, deterministic policy evaluation, named human approvers, an audit trail and immediate revocation.
Why is a prompt or system message not a control?
A prompt expresses intent; it does not enforce a limit. A model can be persuaded, can misread context, and cannot produce an auditable record of its own constraints. Enforcement has to sit outside the model, in code that evaluates each action against explicit rules.
How much autonomy should an agent have?
Start with observation, move to human approval, then to bounded autonomy inside written limits, and widen those limits only on demonstrated behaviour. Each stage should be a deliberate decision with an owner, not a default.
What does a finance or compliance team need to see?
Which agents hold authority, how much of it has been drawn, which actions were refused and why, who approved the exceptions, and how quickly authority can be withdrawn. If those five answers require an engineer, the governance model is incomplete.
Open the demonstration workspaceHow the control model worksSecurity and data handling